THESIS DEFENSE: Equitability and dependence: detecting novel associations in large data sets
, MIT CSAIL
Date: Thursday, May 04, 2017
Time: 10:00 AM to 11:00 AM
Refreshments: 9:45 AM
Location: 32-G449 (Kiva)
Host: Tommi Jaakkola & Josh Tenenbaum, MIT CSAIL
Contact: Teresa Cataldo, firstname.lastname@example.org
Speaker URL: None
TALK: THESIS DEFENSE: Equitability and dependence: detecting novel associations in large data sets
Today's data sets are not only increasingly large but also increasingly high dimensional. Making sense of the wealth of interactions between the large numbers of associated variables in such data is a daunting task, not just due to the sheer number of relationships but also because relationships come in different forms (e.g. linear, exponential, periodic, etc.) and strengths. If you do not already know what kinds of relationships might be interesting, how do you find the most important or potentially unanticipated ones effectively and efficiently? In this talk, I will introduce a property called equitability that enables ranking the relationships among the variables in a data set by strength, independent of type, without requiring any prior knowledge of the forms of relationships for which to search. I will then introduce a statistic that has state-of-the-art equitability in a wide range of settings, discuss how this translates to practical benefits in the search for dependence structure in high-dimensional data, and present an extensive comparative analysis of the landscape of existing measures of dependence. Finally, I will briefly discuss moving from thinking about dependencies to thinking "more causally" about observational data for the purpose of optimal intervention design. Together, these tools empower us to both meaningfully explore and act on our data.
Thesis Supervisor: Prof. Tommi Jaakkola, Prof. Josh Tenenbaum
Thesis Committee: Prof. Tommi Jaakkola, Prof. Josh Tenenbaum, Prof. Stefanie Jegelka
Created by Teresa Cataldo at Thursday, April 27, 2017 at 2:05 PM.